How-to guideNov 19, 2026by Docurensic Team6 min read

How to Tell if an Image Is AI-Generated

There is no single test, and detector scores are guesses. A ranked set of checks that still work, from content credentials to the physics a generator gets wrong, and the question that matters more than 'is it AI?'

How to Tell if an Image Is AI-Generated
In this guide
  1. Key takeaways
  2. Start with where it came from
  3. Then look at the picture
  4. About detectors
  5. The harder problem: real photo, changed
  6. Ask the question that actually matters
  7. Frequently asked questions

Somebody forwards you a photo. It might be a product shot for a marketplace listing, the storm damage on an insurance claim, a profile picture on a job application, or a picture of an event that would be news if it happened. It looks right. It also looks a little too finished, in a way you can't put your finger on.

Two years ago you could count the fingers and move on. That stopped working. The honest position now is that there is no single test, and anyone who offers you one is selling something. What exists is a short stack of checks, some much stronger than others, and the skill is knowing which is which.

Key takeaways

Start with where it came from

Before you look at a single pixel, look for a history.

Check for content credentials. A growing number of cameras, editing programs and image generators attach a signed record to the file saying how it was made. The standard is called C2PA, and the public checker at contentcredentials.org will read it for you. When a credential is present and verifies, it is the best evidence you are going to get, in either direction. Some generators label their own output this way.

When there's nothing there, you have learned nothing. Most cameras still don't sign their photos, and most social platforms and messaging apps strip the record on upload. A missing credential does not mean an image is fake, and it does not mean it is real.

Reverse image search. Run the image through a reverse search. Generated images often trace back to the account that first posted them, sometimes with the prompt in the caption. Real photos trace back to a photographer, an agency or an earlier date. Finding the same scene shot from a second angle by somebody else is strong support that the scene existed.

Read the metadata, lightly. Camera make, lens and capture time suggest a camera. Their absence suggests very little, since screenshots and uploads lose them. We went through what these fields can and can't carry in EXIF data for photo forensics.

Diagram: four kinds of evidence that an image is AI-generated, ranked by strength
Rank the evidence before you count it

Then look at the picture

Current generators are good at surfaces and weaker at consistency. That is where to spend your attention.

Text. Signs, labels, number plates, book spines, the writing on a whiteboard in the background. Headline text is often perfect now. Small incidental text is where things still fall apart, with letters that are nearly letters.

Light. Find the light source. Check that every shadow agrees with it, and that reflections in windows, glasses and water show what should be there. A generator paints plausible light. It doesn't calculate it.

Repetition and logic. Fence posts that drift out of rhythm, a staircase that stops making sense halfway up, earrings that don't match, a watch with no hands, a hand holding a cup at an angle no wrist allows.

Texture. Skin that is flawless everywhere, fabric with no weave, a background crowd where faces dissolve as you zoom in.

Too much polish. Real photographs have dull corners: a bin, a cable, half a stranger. Generated images tend to be composed all the way to the edges.

None of these proves anything alone. A heavily retouched real photo fails several of them. Treat them as reasons to look further.

About detectors

There are plenty of tools that return a percentage. We run one inside our own image analysis, so take this as coming from people with an interest: treat the number as a hint.

These tools learn the statistical habits of particular generators. Shrink the image, screenshot it, pass it through a messaging app, and those habits blur. A detector can be confident and wrong in both directions. It will pass a generated image that was re-compressed twice, and it will flag a real photo that went through a phone's heavy processing, which is most phone photos now. The better tools tell you what they looked at and what they couldn't. The worse ones give you a figure to two decimal places.

A detector score is a reason to do the other checks. It is never the finding.

The harder problem: real photo, changed

In document and claims work, the usual problem is not a picture invented from nothing. It is a real photograph with something altered: damage added to a car, a date changed on a receipt, a face swapped onto an ID. Modern editors use generative fill for this, so the image is partly real and partly made, and "is this AI?" has no clean answer.

For that, the older methods still earn their keep. Compression analysis, noise comparison and clone detection look for a region that doesn't share the history of the rest of the picture. Our walkthrough on telling whether an image has been photoshopped covers them, and they apply whether the patch was painted by a person or by a model.

Ask the question that actually matters

Suppose you could know for certain that a generator made the image. Usually that is not what you needed. A listing photo made by AI is a problem if it shows a flat that doesn't exist, and no problem if it is an illustration and says so. A real, untouched photo is a problem if it shows last year's flood offered as this year's.

So ask what the image is being used to prove, and then verify that. Call the venue. Look up the address. Ask the claimant for three more photos from different angles taken right now, and see how long they take to arrive. Ask for the original file from the phone and not the copy that came through chat. People with a real photo can do these things in minutes.

When the image sits inside a document, such as an ID photo, a stamp or a signature, the file around it can be examined too, and that is usually where the stronger evidence is. See our note on AI-generated fake documents.

Frequently asked questions

Is there a free tool that detects AI images reliably?

No. Free and paid detectors exist and can be useful as a first pass, but none is reliable enough to settle a question alone. Results change when an image is resized, cropped or re-saved.

Do AI images always have mangled hands?

Not any more. Hands, teeth and eyes were weak points in earlier generators and are mostly fixed. Small background text, reflections and physical consistency are better places to look today.

Does missing metadata mean the image is AI-generated?

No. Screenshots, social platforms and messaging apps remove metadata from real photos all the time. Missing metadata is neutral.

What are content credentials?

A signed record attached to an image file describing how it was created and edited, defined by the C2PA standard. When present and valid they are strong evidence. Most images still don't carry them.

Have the image examined

Upload the picture and Docurensic runs error-level analysis, JPEG-ghost, copy-move and noise-floor checks on it, and reads its camera and editor record. Free to start.

Analyze an image free

Keep reading

How-to guideNov 13, 20266 min

How to Find Hidden Text in a PDF

A PDF can carry text nobody reading it will ever see: white on white, under a box, or switched off. How to find it in thirty seconds, and how to tell the innocent kind from the kind written for a machine.